In current-generation computers, the central processing unit (CPU) is connected to the system memory and to peripheral devices by a parallel bus, such as the ubiquitous Peripheral Component Interface (PCI) bus. As data path-widths grow, and clock speeds become faster, however, the parallel bus is becoming too costly and complex to keep up with system demands. In response, the computer industry is moving toward fast, packetized, serial input/output (I/O) bus architectures, in which computing hosts and peripheral are linked by a switching network, commonly referred to as a switching fabric. A number of architectures of this type have been proposed, including "Next Generation I/O" (NGIO) and "Future I/O" (FIO), culminating in the "INFINIBAND.TM." architecture, which has been advanced by a consortium led by a group of industry leaders (including Intel, Sun, Hewlett Packard, IBM, Compaq, Dell and Microsoft). Storage Area Networks (SAN) provide a similar, packetized, serial approach to high-speed storage access, which can also be implemented using an INFINIBAND fabric.
In a parallel bus-based computer system, when a peripheral device needs to deliver data to the CPU, it typically writes the data to the memory over the bus, using direct memory access. When the peripheral has finished writing, it asserts an interrupt to the CPU on one of the interrupt lines of the bus. Bus arbitration ensures that the CPU will not attempt to read the data from the memory until the writing of the data is complete. On the other hand, when the peripheral device and the CPU are connected by a packet-switching fabric, such as an INFINIBAND fabric, they operate asynchronously. Furthermore, the data sent to the memory and the interrupt to the CPU travel over different paths, or channels. Typically, a separate line or channel is provided to connect the interrupt pin of the peripheral device to an interrupt controller of the CPU, bypassing the switching fabric. Therefore, there is no a priori assurance that all of the data will have been written to the memory before the CPU begins reading.
The "race" between the interrupt path and the data path can result in errors (as when a CPU read stalls the data). Care must therefore be taken to synchronize data and interrupt handling and to make sure that the data have been completely written to the memory before the CPU attempts to read it.
A common solution in this situation is to program the CPU to access the peripheral device before accessing the memory, typically by performing a "configuration read" from the peripheral device. In this mode of operation, after the peripheral device has asserted the interrupt to the CPU (indicating that the last item of data has been sent to the memory), the CPU issues a read request through the switching fabric, to read an interrupt cause register in the peripheral device. The peripheral device responds to the read request by sending a packet containing the interrupt cause to the CPU over the same channel as it used to send the data to the memory. Since packets are ordered within a channel, the response to configuration read arrives at the CPU after all of the previous writes have been flushed to memory. The CPU begins to read the data from the memory only after it has received the interrupt cause packet back from the peripheral device. The configuration read thus serves two crucial purposes: it provides the CPU with the cause information that it needs in order to serve the interrupt, and it ensures that the CPU reads the memory only after all of the data have been written there.
This scheme has a number of serious performance drawbacks, however. Every interrupt sent by the peripheral device necessitates an additional exchange of messages through the switching fabric between the CPU and peripheral device. The exchange adds substantial latency--typically 10 microseconds or more--every time the CPU must service an interrupt. Furthermore, since configuration reads are used as synchronization barriers, the CPU is stalled from the moment the configuration read request is issued until its response has arrived. Valuable CPU time is therefore wasted waiting for the interrupt cause to be retrieved.
U.S. Pat. No. 5,689,713, whose disclosure is incorporated herein by reference, describes a method for interrupt request handling in a packet-switched computer system. The system may include a number of interrupt sources, which direct interrupts to any of a number of interrupt handlers. A system controller acts as an intermediary between interrupting devices and "interruptees." It includes an interrupt queue coupled to each interrupt source for receiving multiple interrupt requests, and an output queue coupled to each interrupt handler. The controller thus enables asynchronous data from multiple sources to be conveyed across a packet-switched interconnection, while providing a dedicated channel for interrupts associated with the data packets.